Your AI Capacity Plan Needs a Date and a Workload

Texas had more than 474 gigawatts of large-load requests in its interconnection queue by August, but fewer than 9,000 megawatts had approval to energize, according to an October 2 contributed article in TNW. That is roughly two percent of requested capacity reaching the approval stage. The rest is a very impressive number to put in a presentation whose audience has not yet asked when the machines turn on.

The gap matters, but it is not a measurement of unmet operating demand. The account describes applicants filing the same project across multiple sites and territories; requested capacity includes options that will not all become facilities. Approval to energize also does not establish actual consumption. For a CTO, the useful conclusion is narrower: the queue cannot certify the capacity behind your deployment schedule. Adding up requests tells you how much people have asked for, with considerable enthusiasm and some duplication.

The article proposes a more demanding test: a contracted quantity, a named counterparty and a delivery date. Those details turn a site’s proximity to infrastructure into a proposition someone can evaluate. A transmission line passing the property is encouraging geography; it does not answer what your project is entitled to receive or when. The distinction is especially concrete for gas supply, where the account identifies transportation agreements, delivery points and firm versus interruptible service as consequential terms. TNW published this as a contribution rather than newsroom reporting, but its contract test is a useful procurement question: which document makes the capacity usable on your calendar?

A delayed connection can do more than move revenue into a later spreadsheet column. In the same TNW account, GoodPeak CFO Vicente Garrido says missing a project gate can move it into a different study cohort, change assumptions or required upgrades, and sometimes create contractual termination rights. He represents a developer participating in this market, so this is an informed participant’s account rather than a neutral audit. The mechanism nevertheless deserves attention: the schedule can change the terms under which the project remains viable.

If an anchor customer can leave after a delivery deadline slips, the project may lose the commitment that made its capacity commercially valuable. That is a different planning problem from applying a discount rate to otherwise unchanged future cash flows. The article also reports that Texas legislation and subsequent reforms made queue positions more expensive to hold through $50,000 per megawatt of financial security, verified site control and disclosure requirements. At one gigawatt, that is $50 million committed before construction. These requirements make speculative applications more costly; they do not themselves establish a reliable completion date. TNW

Nor does this justify concluding that America has plenty of generation and merely needs better paperwork. The article cites Ascend Analytics’ estimate that more than 80 percent of new large loads seeking ERCOT interconnection will lack matching generation online by 2030, alongside uncertainty about how long new processing regimes will take. The practical implication is to examine what happens when a date moves: which studies change, which upgrades become necessary, and which commitments can disappear. A deployment plan that records only the expected energization date has documented the happy path. Engineering organizations already know how expensive that particular form of optimism can become.

There is a complementary capacity question inside the machine room: what workload can the chosen model serve on the hardware you intend to operate? Aleph Alpha’s technical account of Kolibri offers a concrete design decision. The company says a 123B-parameter configuration could handle only three 256k-token user queries on two H100s, while its chosen 78B configuration handled eighteen concurrent requests and decoded 28 percent faster. Those are vendor-reported results, but they explain a trade-off worth testing. A larger model improved performance in the company’s experiments; serving costs helped determine where it stopped scaling. Aleph Alpha

The architecture gives that decision technical substance. Kolibri has 78B total parameters with approximately 3B active, and forty of its fifty layers use a focused 512-token attention window while ten process the full context. Aleph Alpha says the restricted windows bound decode computation and memory in those forty layers regardless of context length. That is more informative than waving the active-parameter count at a capacity spreadsheet and hoping the remaining configuration politely disappears. The serving comparison concerns 256k-token queries; the model’s advertised support for contexts up to one million tokens does not establish eighteen-request concurrency at that length. Aleph Alpha

For practitioners, this suggests a paired acceptance test. Ask the infrastructure supplier for the contracted quantity, counterparty and energization date described in TNW; test the model’s concurrency, latency and answer quality on the intended workload. Kolibri supplies a candidate engineering lever, not an established energy saving or a remedy for Texas interconnection constraints. The two decisions meet in the deployment plan: deliverable power determines whether the site can operate on commercially usable terms, while measured serving performance determines what that operation can support. The opening queue number and the model’s parameter count each omit something essential. Your launch date needs the omitted details.